Google DeepMind open-sources WeatherNext for cyclone forecasts

Google DeepMind and Google Research have open-sourced WeatherNext models after reporting a major improvement in tropical cyclone forecasting in Nature. The system predicts a storm’s track, intensity and wind structure in one AI model, instead of separating global path forecasting from more local intensity modeling.

Google says WeatherNext Cyclones gives forecasters more than a full day of additional useful lead time on average: its three-day forecasts match what previous models could provide at two days. The company frames that as roughly a decade of progress compressed into one model.

The release includes WeatherNext Cyclones, WeatherNext 2 and WeatherNext 2-mini, a smaller version that can run in a public Colab notebook. The models can generate 15-day forecasts and large ensembles of possible storm scenarios, which helps forecasters reason about low-probability but dangerous outcomes such as rapid intensification.

This is not a replacement for official meteorological warnings. But as an open research tool, it gives weather agencies, researchers and nonprofits a stronger base for local forecasting and disaster preparation.